centrality of the supply chain networkfaculty.baruch.cuny.edu/lwu/papers/splc_ov3.pdf · 2004 2014...

Centrality of the Supply Chain Network

Liuren Wu

Baruch College

December 7, 2016

Wu (Baruch) Supply Chain Network Centrality 12/7/2016 1 / 25

Statistical estimation v. structural identification

One of the central themes of research is to estimate/test various relations.

There are two broad approaches:

1 Statistical estimation

Regressions, dynamic statistical relation specifications/estimations(VAR, GARCH, DCC,...)

PCA, Trees, neutral network, boosting, ... (Google translate)

More data and more computing power → better statistical learning

2 Structural identification

Industry classification: Companies within the same industry tend tobehave similarly and tend to move together

Linguistic structures: Translation based on understanding (grammar) ofthe two languages

Supply chain relation: A shock to the supplier can generate rippleeffects via the supply chain.

One can gain a better understanding of the relations by going back andforth between the two approaches.


What can we learn from the supply chain information?

If we know that A is the supplier of B, then what?

Commonly formulated questions in the supply chain (finance) literature:

What happens to the customers (and the customer’s customers) if asupplier company goes under?

What does a good performance of a supplier company foretell about itscustomers?

Mostly focus on the lead-lag relations between supplier-customer pairs(stock returns, earnings etc)

Statistical analysis of the supply chain:

Identify pairs, and verify whether the relation is as expected.


But it’s a chain (network), not a stand-alone pair

A is the supplier of B, C, D, the customer of E, F, ... Pairs do not exist in isolation

What can we learn, what do we want to learn, from this network?


Questions on the network

Are there suppliers that are more important to the economy than others? Inwhat sense/metric?

Shocks to which company generate the strongest ripple effect? Why?

Which companies are the most central to the network?

How do we define the relative importance of a company within thisnetwork?

What roles do these central companies play in the aggregate economy?

What types of companies tend to be the central ones in the network?

How do they behave differently?

Do changes in these central companies foretell the transition of thebusiness cycle, or structural changes in the economy? How?

What I try to explore in this paper:

Explore metrics of relative importance (centrality) for each companywithin the supply chain network.

Explore what types of companies tend to be “central” and how“central” companies interact with the aggregate economy.


The supply chain data

Historical firm-level supplier-customer relationship information is fromFactSet Revere.

The SEC has a mandate (rule SFAS 131) that calls for companies todisclose their customers if their revenue exposure is 10% or greater.

Some companies choose to disclose more customer and supplierinformation if they believe the information will help improve theirattractiveness and draw more attention from investors.

FactSet collects the information from the company’s regulatorydisclosure reports, annual reports, and other primary sources.

I confine the analysis to constituents of the S&P Composite 1500 index.

From 2004 to 2014, I take snapshots of the database at the beginning ofeach year, and merge the information with price and accounting data fromBloomberg.

One can sample the information more frequently, but not sure aboutthe practical feasibility of more frequent information updates.


The supplier network matrix

Based on the FactSet data, I construct a supplier network matrix {Ai,j}Ni=1

at the beginning of each year on the S&P 1500 companies

Ai,j = 1 if company i is the supplier of company j ; 0 otherwise.

Each column j represents an index of suppliers for company j .

Each row i represents an index of customers that supplier i supplies to.

The matrix does not distinguish the relative significance of each supplier.

FactSet database has some sales exposure information, but is far toosparse to be useful.

Other metrics: irreplaceability, product type, value added, profitmargin, sales...

The matrix is incomplete, and has potential biases:

A S&P1500 company can have important suppliers outside the 1500list — pruned network.

There are potentially unidentified suppliers within the 1500 list.

FactSet uses cross-validation, which may create some bias/noise— A may list B as an important customer, but A may not be asignificant supplier to B.


Understand the topology of the supplier network matrix

The supply chain network is a directed network, similar tothe hyperlink network...

One can draw an arrow from a customer to each of its suppliers.

The network matrix A is positive, but not symmetric.

A Peers Network would be an undirected network, with a positivesymmetric network matrix.

The network literature has proposed a long list of centrality measures todefine the relative importance of a vertex (a company) within the wholenetwork.

I adapt some of these centrality concepts as measures of the relativeimportance of a supplier to the overall economy.

1 Degree centrality2 Eigenvector centrality3 PageRank4 Kleinberg supplier authority v customer hub centrality

I analyze the relevance and applicable scenarios of each measure.


Degree centrality: The number of direct customers

Intuitively, a supplier company is more important to the economy if it is thesupplier of many companies.

A simple measure of centrality for a supplier, the degree centrality, can bedefined as the number of companies that this company supplies to.

From the perspective of the network topology, if one draws an arrow from acompany to each of its suppliers, the degree centrality measures the degreeof a vertex, or the number of arrows pointed to a particular supplier.

From the supplier matrix A, one can compute the degree centrality vector, c,as the simple sum of each row,

ci =∑j

Ai,j , or in matrix form, c = Ae (1)

where e denotes a vector of ones.

All centrality measures are relative: One can normalize the metric to require,for example, e>c = N or e>c = 1.


Top 10 suppliers with the most customers

2004

GE HPQ IBM INFA INTC IWOV MSFT ORCL ROK WBSN0

10

20

30

40

50

60

70

Numb

er of cu

stome

rs

2014

BFAM BRK/B CR GE IBM MSCC MSFT NPO ORCL SXI0

10

20

30

40

50

Numb

er of

custo

mers

Top suppliers have 30-70 customersWu (Baruch) Supply Chain Network Centrality 12/7/2016 10 / 25

Eigenvector centrality: The importance of customers assuppliers

Each customer can be a supplier to other companies. Its importanceincreases if it is an important supplier itself to many other companies.

Shocks to a supplier are more likely to cause a stronger propagation(ripple effect) if it is the supplier of important suppliers.

Eigenvector centrality gives each supplier a score proportional to the sumof the scores of its customers:

ci = (1/λ)∑j

Ai,jcj . (2)

N equations, (N + 1) unknowns, up to a scale.

In matrix notation, we have, λc = Ac.

The solution is proportional to the right leading eigenvector of matrix A.

Since all elements of A are non-negative, the Perron-Frobenius theoremstates that the leading eigenvector has strictly positive components.


PageRank: Discount customers with many suppliers

A customer of many suppliers contributes less to each of these suppliers

If one company has a high eigenvector centrality, all its suppliers willhave high eigenvector centrality.

In many cases it means less if a company is only one of the manysuppliers (esp. if they are substitutes).

PageRank centrality discounts the centrality of each customer by itsnumber of suppliers.

ci = (1/λ)∑j

Ai,jcjkj, kj = max

(1,∑i

Ai,j

). (3)

Brin and Page (1998) proposed the idea in their modeling and rankingof the internet hyperlinks.

Can be solved as the leading eigenvector of a scaled matrix:λc = AK−1c.


Kleinberg centrality

Suppliers and customers can be important for different reasons.

A customer is important to the economy if it needs many differentparts supplied by different suppliers and hence acts as a hub ofdifferent (non-substitute) suppliers.

This goes to the opposite direction of PageRank discounting, fordifferent reasoning.

A supplier is important if it supplies to many such important hubs.

Kleinberg (1999) proposes to construct two types of centralities, labeled asthe authority (supplier) centrality and the hub (customer) centrality, toquantify each vertex’s prominence in the two roles.

Define supplier centrality s as the sum of the customer importance, andcustomer centrality c as the sum of its supplier importance:

si = α∑j

Ai,jcj , cj = β∑i

Ai,jsi . (4)

Combining the two and in matrix notation, AA>s = λs, A>Ac = λc.The solutions are the leading eigenvectors of AA> and A>A, resp.


Differences among the four centrality measures

Supplier importance is the average importance of its customers.

The four centrality measures differ in the customer importance definition:

1 Degree centrality: All customers are equally important.

2 Eigenvector centrality: Important suppliers are important customers:customer importance = supplier importance.

3 PageRank centrality: Discounts the importance of a customer by itsnumber of suppliers:customer importance = supplier importance/Number of suppliers.

4 Kleinberg centrality: A customer is more important if it has moreimportant suppliers:customer importance=sum of the importance of its suppliers


Numerical issues with eigenvector centrality

Supplier centrality avoids degenerations in Eigenvector centrality:

If a company is not the supplier to any company, it will lead to a rowof zeros in the supplier matrix. The eigenvector centrality is zero.

Imagine another company, which is a supplier to thousands of suchcompanies — The eigenvector centrality will remain zero.

This can go on and on...Supplier centrality avoids this issue via adifferent definition of customer centrality.

In practical implementation, PageRank employs an ad hoc fix: Add asmall quantity to all elements of c: c = αAc + βe.

β can also differ from company to company and can be used as a priorobtained from other sources of information (e.g., sales, size, etc)

Adding how much is a bit arbitrary.

The separate definitions of supplier/customer importance probably makemore sense in most practical situations.

It would be even better if one can treat substitute suppliers differentlyfrom non-substitute suppliers.


Distributional behavior of centrality estimates

2004 2014

0 200 400 600 800 10000

20

40

60

80

100

120

Cu

mu

lativ

e c

en

tra

lity

con

trib

utio

n,

%

DegreeEigen

AuthorityHub

0 200 400 600 800 10000

20

40

60

80

100

120

Cu

mu

lativ

e c

en

tra

lity

con

trib

utio

n,

%

DegreeEigen

AuthorityHub

Degree centrality is most uniform; eigenvector centrality is mostconcentrated.

Important customers are more concentrated than important suppliers.


Average rank correlation between centrality measures

Degree Eigenvector Authority HubDegree 1.000 0.776 0.891 -0.086Eigenvector 0.776 1.000 0.805 -0.045Authority 0.891 0.805 1.000 -0.111Hub -0.086 -0.045 -0.111 1.000

The three supplier centrality measures are highly correlated — They capturesimilar behavior.

Supplier and customer centralities show small negative correlation.

This approach can indeed separate two types of companies.


Central company characteristics and behaviors

Centrality TR ER Beta IV RV CorrQuantile Panel C. Supplier authority centrality

1 0.131 0.026 1.223 0.352 0.319 0.5402 0.117 0.032 1.262 0.378 0.330 0.5153 0.122 0.021 1.285 0.396 0.342 0.4964 0.121 0.041 1.274 0.381 0.335 0.4995 0.108 0.022 1.338 0.391 0.350 0.511

Panel D. Customer hub centrality1 0.123 0.060 1.240 0.388 0.334 0.5032 0.112 0.031 1.210 0.368 0.337 0.5063 0.096 0.028 1.251 0.365 0.325 0.5174 0.105 0.006 1.243 0.365 0.328 0.5155 0.112 0.020 1.193 0.344 0.312 0.532

Central authorities (suppliers) tend to have higher volatility, but lowercorrelation with market.

Central hubs tend to have lower volatility, but higher correlation with market.

Highly connected firms tend to have lower returns?


Top 10 suppliers with the most number of customers

2004 GE HPQ IBM INFA INTC IWOV MSFT ORCL ROK WBSN2005 A ACN CNQR GE HPQ IBM INTC IWOV MSFT ORCL2006 ACN GE HPQ IBM INFA IWOV MANH MSFT ORCL ROVI2007 EMC GE HPQ IBM IWOV MSFT ORCL PRGS QTM ROVI2008 DDR EMC HPQ IBM IWOV JDAS MSFT ORCL PRGS ROVI2009 DDR EMC HPQ IBM JDAS LXP MSFT ORCL PRGS ROVI2010 DDR EMC HPQ IBM JDAS LXP MSFT ORCL PRGS ROVI2011 ACN DIS FRT HPQ IBM INTC JDSU LXP MSFT PTC2012 ADC BFS EP FRT HPQ IBM INTC LFUS MSFT PWR2013 BRK/B CR HON HPQ IBM IRC LFUS MSFT NPO ORCL2014 BFAM BRK/B CR GE IBM MSCC MSFT NPO ORCL SXI

Mostly tech companies in earlier years, some industrial goods later, and oneservice.


Top 10 suppliers with the highest eigenvector centralities

2004 CYMI IBM INTC KEM NTAP ORCL PSEM SMTC TTMI XLNX2005 A AMCC CYMI IBM IRF KEM NTAP PSEM SMTC TTMI2006 A ATML IBM IRF MSFT ORCL PSEM ROVI TTMI XLNX2007 ATML BRKS CA IBM JBL JDSU KEM PSEM QTM SNIC2008 ATML AVT COGT CYMI IBM KEM ORCL PSEM ROVI SNIC2009 ATML CYMI EMC HPQ IBM JAVA MSFT QLGC ROVI SNIC2010 ATML EMC HPQ IBM JAVA MSFT QLGC QTM ROVI SNIC2011 ADBE ATML CDNS HPQ IBM INTC KEM MSFT QTM SNIC2012 ATML CDNS HPQ IBM INTC IRF KEM LFUS XCRA MSFT2013 IBM INTC IRF JBL KEM LFUS LSI NVDA ORCL XLNX2014 AKAM COHU EXAR IBM INTC LFUS LSI MMM NVDA XLNX

Mostly tech companies.


Top 10 companies with the highest supplier authoritycentralities

2004 IBM INFA INTC IRF KEM MSFT ORCL SANM WBSN XLNX2005 A EMC GE HPQ IBM IRF KEM MSFT ORCL SANM2006 A HPQ IBM INFA IRF IWOV MSFT ORCL PRGS ROVI2007 ATML EQIX IBM IRF IWOV JCI MSFT PRGS QTM ROVI2008 BCSI DDR EQY FICO FRT IWOV MAC PEI PRGS ROVI2009 DDR EMC EQY HPQ IBM MAC MSFT PEI PTC PRGS2010 DDR EMC EQY HPQ IBM KRG MAC MSFT PEI PRGS2011 DDR FRT HPQ IBM JAH KRG MAC MSFT PEI SPG2012 ADC AKR DDR EQY FRT KIM KRG MAC PEI SPG2013 ADC CR DDR FRT HON IRC KIM KRG LEG MAC2014 ADC CR DDR FRT IRC KRG LEG MAC MMM SXI

Mostly tech earlier years. 2014 has 6 financials (real estate), 0 tech.


Top 10 companies with the highest customer hubcentralities

2004 AVT BA CSCO DELL GE HPQ IBM INTC LMT WMT2005 BA CSCO DELL F GE HPQ IBM INTC LMT WMT2006 BA CSCO F GE HPQ IBM INTC LMT MSFT WMT2007 BA CSCO F GE HPQ IBM LMT SHLD T WMT2008 BA CSCO HD HPQ IBM LMT SHLD T TGT WMT2009 BA HD HPQ IBM KSS MSFT SHLD T TGT WMT2010 BA BBY HD HPQ IBM KSS SHLD T TGT WMT2011 BBY HD HPQ IBM JCP KSS M SHLD TGT WMT2012 BBY COST HD JCP KSS LOW M SHLD TGT WMT2013 BA GE HD KSS LMT LOW NOC SHLD TGT WMT2014 BA BBY COST HD LMT LOW NOC SHLD TGT WMT

Many stores: BestBuy, Costco, Home depot, Lowe’s, Sears,Target, Walmart


Statistical behaviors of central stock portfolios

Portfolios are formed equal weighting with top 10 central companies:

Quarterly total return Quarterly excess returnYear β ρ 3-month 6-month 3-month 6-monthDegree 1.075 0.879 0.177 0.043 0.088 0.015Eigenvector 1.245 0.795 0.185 0.079 0.115 0.093Authority 1.208 0.840 0.232 0.103 0.232 0.177Hub 0.984 0.872 0.122 0.016 0.010 -0.002

β: Regression against the S&P 500 index: Central supplier portfolios seemto be more volatile.

ρ: correlation with the index

Predictive correlations with the index over 3/6-month horizons:Excess returns on central supplier portfolios predict future stock returns.Excess returns on central customer portfolios do not.


Concluding remarks and future works

It should be useful to look into the topology of the supply chain network tobetter understand its aggregate characteristics.

The performance of important suppliers lead the economic cycle.

Other unexplored questions:

Sector rotation in the business cycleAre important suppliers “too central to fail”?

The topology should also be useful for firm-level analysis

A shock to any companies first propagates to all its customersthen to all the customers of its customers, ...The impulse-response can be analyzed through the network matrix Aand its powers, A2, A4, ...


The devil is in the details

Much work can be done on refining the accuracy of the network matrix

Expand the matrix to be less pruned — Too much pruning can distortthe topology...

Accommodate/Include other informations (such as fraction ofsales/COGS) to refine relative significance of each customer/supplier

Use product-level information to differentiate substitute suppliers w.non-substitute suppliers

Combine network centrality with other centrality/importance measures:Katz centrality: c = αAc + βc0,with c0 being the prior constructed from other criteria.


centrality of the supply chain networkfaculty.baruch.cuny.edu/lwu/papers/splc_ov3.pdf · 2004 2014...

Documents